Tuesday, August 27th 2024

AMD Threadripper 9000 "Shimada Peak" with 96 Zen 5 Cores Spotted

A shipping manifest has revealed a new 96-core Ryzen Threadripper CPU codenamed "Shimada Peak." This processor is expected to incorporate Zen 5 CPU cores and maintain a configuration similar to AMD's current Zen 4 Threadripper flagship. The new CPU will likely be compatible with existing DDR5 HEDT motherboard platforms like TRX50 and WRX90. Its architecture appears to mirror that of AMD's previous 96-core Threadripper and EPYC models, featuring twelve 8-core CCDs and a single IO die. These similarities suggest that the primary change lies in the CCDs.

The Shimada Peak series is unlikely to debut in the near future. Given that the previous generation launched in October 2023 after a considerable delay, a similar timeline may apply here. AMD has remained silent regarding its Threadripper roadmap. Reports suggest AMD is preparing to unveil multiple products soon, the lineup is said to include the Ryzen AI 300 PRO, EPYC Turin, and Instinct MI325X. Following this, the next major showcase is expected to be CES. At this event, Kraken or Strix Halo are anticipated to make an appearance.

While Zen 5's gaming performance improvements have been modest, they have also shown significant gains in workstation and server applications. Phoronix benchmarks and our reviews highlight substantial performance boosts in Machine Learning, Database, HPC, and cryptography workloads compared to Zen 4. These improvements make Zen 5 Threadripper and EPYC CPUs appealing to specialized users.
Source: OC3D
Add your own comment

28 Comments on AMD Threadripper 9000 "Shimada Peak" with 96 Zen 5 Cores Spotted

#26
efikkan
AssimilatorThere has to be a middle ground between "actual server grade" and "consumer". People who want a lot of IO shouldn't be shut out.
Indeed.
And it shouldn't be hard for AMD to bring some of their next gen Threadrippers quickly either, and since it's not a new platform, they could easily launch a partial lineup if yields are an issue. A smaller HEDT socket would be preferable in the long term, but in the meantime I wish they did something like the following to appeal to the forgotten "semi-pro" and professional markets;
- Offer 12 and 16 core models, below $1000.
- Strongly consider an 8 core model too, golden sample with good clocks, ~$500. (I'm pretty sure this would still sell)
- Ask motherboard partners to offer more basic motherboards in the ~$600 range.
(Additionally, make workstation motherboards have the CPU socket in the "right" orientation for air-cooling in cases, not in the "rack orientation" (which is fine for server boards).)
This would likely increase the sales of the platform drastically.
Posted on Reply
#27
kapone32
AssimilatorLast time I checked I didn't have to dick around with Microsoft Game Bar, running in Admin mode, and/or specific Windows versions to get games to run properly on my 3960X. Based on that I'd say TR is better designed to game on than Zen 5.

I didn't buy my current system because I wanted or needed 24 cores, I bought it because I needed 32 lanes of PCIe which is simply not an option on AMD's so-called consumer platform. That's literally the only reason - the IO. I didn't particularly want a 3960X with its age and heat output and ridiculous price tag, but AMD literally gave me no other choice except to go fuck myself. I'd have been perfectly satisfied with 8 cores but nope, if you want anything more than the meagre 28 lanes of PCIe that their so-called consumer platform puts out you have to buy HEDT.

Meanwhile back in socket 939 days, 32+ lanes of PCIe was the norm even on entry-level boards.
Gaming was fine on Threadripper. The issue was cost. You could get a 1900X for $200 when X399 was new and those MBs were like today's boards in price but so much better in flexibility. The issue is that the 5950x blew the 2920X away in CPU performance so if you wanted that you had to sacrifice PCIe joy and go back to AM4. I lament that I cannot justify buying TR today but I do enjoy the special effects in Movies that TR introduced.
Posted on Reply
#28
efikkan
kapone32Gaming was fine on Threadripper. The issue was cost. You could get a 1900X for $200 when X399 was new and those MBs were like today's boards in price but so much better in flexibility. The issue is that the 5950x blew the 2920X away in CPU performance so if you wanted that you had to sacrifice PCIe joy and go back to AM4.
Cost was a large part of the problem, another was the availability. The Zen 3 Threadrippers (59xxWX series) was initially OEM only, it took a while before they showed up in retail channels, and even that wasn't available in every country. It also released very close to Zen 4. They need to get their act together and release these together with their Epyc counterparts.

Another aspect is the difficulty of benchmarking "workstation workloads" in a fair and representative manner.
Benchmarking games or large batch server workloads are by comparison "easy" to perform, while a typical workstation is used for mixed workloads. Whether we are talking about "power users"/prosumers or actual professionals (small business owners or in a corporate setting), and whether the workloads consists of development, CAD, modelling, graphics, video, etc., they pretty much all run many different applications at the same time. Not with high load, of course, but different applications "competing" with medium loads are more typical. So what people usually need is something with a decent mix of (consistent) CPU speed, memory bandwidth and PCIe.

I can probably count on a single hand the amount of "professionals" I've seen who only run a single application at the time, and then close it when switching between tasks, the more typical use case is a whole lot of applications open and users switching between tasks during the day. Even for developers (which is my area), there is typically multiple development/debugging tools, a web browser with 100 tabs and a couple of VMs etc. running simultaneously, closing and opening tools would be very unproductive, so a pure code compilation benchmark doesn't reflect the actual usage.
And I believe for those heavily into video editing, I don't think those spend their time watching their videos encode, and if they have the volume for it, they have a dedicated server for encoding and a workstation for actual productive work, so once again the typical benchmark doesn't necessarily reflect the actual workflow.

So I think part of the "problem" which makes the 16 core mainstream CPUs seem more attractive than they in reality are is the choice of benchmarks. Not only is there a focus on synthetic benchmarks (which are pointless for real world usage), but also the weighting of either batch jobs and/or workloads that are too niche. One good example of this is Cinebench, which is probably the most featured benchmark in recent years, and while it is a real workload, it's only relevant for users of Cinema4D. This also becomes a problem if people use aggregate benchmarks to pick workstation parts, as they might actually pick an inferior part (or miss out on the advantage of a better one).

But I don't have the solution of how to benchmark workstation workloads more fairly either. It's a tricky one. Despite not having used Threadrippers, I can attest to the night and day difference in user experience of working on Intel's HEDT platforms (x299 and x79) vs. their mainstream contemporaries, and I have every reason to believe Threadripper would have similar benefits. ;)
kapone32I lament that I cannot justify buying TR today but I do enjoy the special effects in Movies that TR introduced.
While I can't speak for your usage, if anyone keeps upgrading systems due to either memory or IO limitations, going for a Threadripper (or Xeon-W) might actually aid in the system's longevity. Not to mention workstation motherboards in generally are better designed.
Posted on Reply
Add your own comment
Nov 21st, 2024 08:26 EST change timezone

New Forum Posts

Popular Reviews

Controversial News Posts